A K-space guided invertible diffusion model for accelerating MRI reconstruction.
Authors
Affiliations (1)
Affiliations (1)
- School of Computer, Jiangsu University of Science and Technology, Zhenjiang, China.
Abstract
Magnetic resonance imaging (MRI), a core technology in clinical medical diagnosis, has long aimed to balance imaging quality and acquisition efficiency. Traditional MRI techniques, limited by the Nyquist-Shannon sampling theorem, are faced with significant challenges in high-acceleration-factor imaging, including artifacts and detail loss due to sparse sampling. While existing parallel imaging, compressed sensing, and deep learning methods have improved reconstruction efficiency, they have limitations such as insufficient integration of electromagnetic physics models with biological tissue prior information, inadequate high-frequency information recovery, and high computational complexity. To address these persistent limitations in high-acceleration MRI reconstruction, this study proposes a K-space guided invertible diffusion model (KG-IDM) designed to improve reconstruction fidelity. KG-IDM enables high-quality MRI reconstruction through Cross-domain Feature Injection (CFI) with Adaptive Channel-Spatial Interaction Attention (ACSIA) for multi-source feature alignment, and a K-space Guided Bimodal Spectral Enhancement Network (KG-BSEN) for dual-path frequency-spatial enhancement. This framework jointly exploits spatial and spectral representations, enhancing cross-domain feature consistency and reducing reconstruction artifacts. Experiments on the fastMRI single-coil knee dataset using retrospectively undersampled complex-valued k-space at <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>4</mn> <mo>×</mo></mrow> </math> , <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>6</mn> <mo>×</mo></mrow> </math> , <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>8</mn> <mo>×</mo></mrow> </math> , and <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>10</mn> <mo>×</mo></mrow> </math> acceleration show that KG-IDM achieves the highest mean PSNR and SSIM and the lowest mean NMSE relative to all nine baseline methods. At <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>4</mn> <mo>×</mo></mrow> </math> acceleration, KG-IDM achieves <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>32.05</mn></mrow> </math> dB PSNR, <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>81.35</mn> <mo>%</mo></mrow> </math> SSIM, and <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>0.0230</mn></mrow> </math> NMSE. At <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>10</mn> <mo>×</mo></mrow> </math> , it achieves <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>25.58</mn></mrow> </math> dB, <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>62.55</mn> <mo>%</mo></mrow> </math> , and <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>0.0815</mn></mrow> </math> , respectively. Paired volume-level Wilcoxon tests over 199 evaluation volumes showed statistically significant favorable differences in all 108 comparisons, with all Holm-adjusted <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>p</mi> <mo><</mo> <msup><mn>10</mn> <mrow><mo>-</mo> <mn>4</mn></mrow> </msup> </mrow> </math> . By integrating measurement-guided cross-domain feature fusion with frequency-spatial refinement, KG-IDM consistently improves reconstruction fidelity across the evaluated acceleration factors and provides a promising framework for accelerated MRI reconstruction.